On Sunday I was stuck in a nightmare scenario. It was the Sunday after Thanksgiving, and all of California was heading home from visiting family over the holidays. What’s normally a 5 and half hour drive between Los Angeles and Silicon Valley became a nine and a half hour drive, with much of the day spent crawling forward in stop-and-go traffic at five miles per hour. I’m almost certain that I would have completely lost my mind if I didn’t have Tesla Autopilot and the “Full Self-Driving” package.

Despite being in the car for 9.5 hours by myself, I hardly touched the steering wheel or pedals all day. Instead, my Tesla Model 3’s software drove me all the way across California, completely automated. Autopilot today requires human supervision, and the supervising human will often have to take over control of driving the car if the computer makes a mistake. In this case however, Autopilot was able to drive for nearly 10 hours without any sort of takeover at all, except for one 30 minute stop to charge at the new Kettleman CIty, California superchargers.

So what?

So what’s the big deal about Autopilot driving all the way from Los Angeles to Silicon Valley with zero takeovers? After all, Autopilot has been able to do this for a long time. We ourselves posted a video recording of Autopilot + FSD Beta driving from Los Angeles to Silicon Valley with zero takeovers required all the way back in January. As if that wasn’t enough, we also recorded Autopilot crossing California with zero takeovers in August. So is doing it for the third time with the latest software really news?

I would argue yes. This drive was significantly harder than any of the previous drives along the same route. I remember when Navigate on Autopilot was first released, and how much it struggled with dense traffic. Sunday was some of the densest most painful traffic I’ve ever seen in my life, and Pure Vision Autopilot handled it like a charm. At 9.5 hours, this drive lasted four hours longer than some of the previous attempts. That’s a lot more opportunity for something to go wrong and require a takeover, and a lot more impressive to pull off. Autopilot was not always able to handle intense traffic conditions like these, and the fact that it can now is extremely impressive. If I had to handle stop-and-go traffic for almost ten hours, I would be dealing with PTSD right now instead of writing a blog post. Thank heaven for Autopilot.

Pure Vision

Another difference between this third zero takeover cross California drive and the first recording in January is that Autopilot is now running on a new “Pure Vision” codebase. This means that the radar that was previously used to help drive the car is now disabled and the car is now getting all of its information visually from cameras, the same way humans drive.

When I first heard about the plan to ditch radar and rely solely on computer vision to judge distance and drive the car, I thought Elon Musk might have lost his mind. I was wrong. Pure Vision is a crowning achievement of the Tesla Autopilot team, and a proof of concept that deep learning can attain the level of reliability needed to drive a car.

In almost ten hours of driving, the vision system never failed to keep a safe distance between my Tesla and the car it was following. In the early days of Autopilot when Tesla was using automotive radar to keep a safe gap between the car in front, I remember sometimes having to disengage Autopilot because a car had stopped suddenly and Autopilot wasn’t slowing down fast enough. That doesn’t happen anymore. Just by looking at the cars, trucks, and other vehicles ahead, Autopilot was able to keep a safe distance and never made a serious misjudgment requiring takeover. This was a very challenging scenario encompassing everything from driving 80 miles an hour, to stop and go traffic at 5 miles per hour, to sudden stops from 80 to 0 and sudden accelerations from 0 to 80. This was a true torture test for pure vision Autopilot and the system passed with flying colors.

Properly judging the distance to an object always seemed like one of the hardest tasks to do accurately with computer vision. In fact, not too long ago the general consensus was that it just couldn’t be done without expensive laser scanners bolted all around the car. If Tesla can make Pure Vision’s distance calculations this flawless, there’s no reason why they can’t make every computer vision task Autopilot has to perform work just as reliably.

Now, pure vision isn’t perfect. In some ways, I felt that the old radar codebase handled stop-and-go traffic much more comfortably for passengers than pure vision. However, the benefits of pure vision are now clear and I feel confident that the new Pure Vision Autopilot will surpass the old code in stop and go traffic and any other scenario in the coming months and years.

I Expected It

The first two times I recorded FSD driving from Los Angeles to Silicon Valley with zero takeovers, I was shocked. I couldn’t believe the software had done it, because I had never seen it before. This time was different — throughout the drive I found myself actually expecting that the car would arrive in Silicon Valley without any human takeovers.

I had to remind myself a few times not to count my chickens before they hatched, telling myself that this was a very challenging drive where many things could go wrong. But somehow, I knew it would work. The key for Autopilot isn’t just being able to drive from Los Angeles to the Bay Area without human control — to achieve its true potential, Autopilot needs to be able to pull this off every time a user asks it to try.

It’s hardly a scientific metric, but the fact that I had a feeling the drive would go off perfectly before it did says something too. Don’t get me wrong, the software is nowhere close to being able to do this trip on its own 100% of the time. But it has clearly gotten a lot better at completing this drive itself since January, and that’s pretty remarkable.

Millions of Robotaxis

The biggest and most obvious takeaway of all this is that it’s only a matter of time before there are automated vehicles that can take us between Los Angeles and Silicon Valley automatically any time we want. These vehicles will allow us to sleep, use our phones, or do anything else we’d like to while the software handles all the driving. What’s more, this software will be much safer and more reliable than humans, crashing much less frequently. People are so desensitized to a decade of autonomy hype that few are paying attention to the transformation that’s happening now with Autopilot. As this software is perfected, millions of Teslas on the road today will be able to run it. All Tesla has to do is push out an improved version, and millions of Robotaxis wake up overnight.

These won’t be restricted and limited Robotaxis like the ones Waymo is running in Arizona. You’ll be able to go anywhere anytime, and even keep these cars in your garage. It’s going to completely transform everything about the way we live.

“But there are so many problems!”

Many people who are skeptical or pessimistic about FSD point to numerous problems and annoyances people are encountering running the beta. It is easy to point to an annoying issue, like the car braking when it doesn’t need to, and say that the software will never be ready for the masses. And you know what? There’s a lot of truth to that. The software is still in beta, and in many ways is not ready for the masses yet.

But just think about this: When I first got my Model 3, all it could do was basic lane-keeping on city streets. It couldn’t automatically change lanes until a software update enabled it. Features like summon didn’t come on the car when I first bought it, they were added over the air. Navigate on Autopilot, Smart Summon, Autopark, FSD Beta… I can’t even count all the improvements to Autopilot that have happened over the years. The software has just gotten better and better, while also becoming safer and safer. It has metamorphized from a commodity lane-keeping system found on many vehicles to something that can now do entire drives from parking lot to parking lot without any human input at all in some cases. It has thousands of real users like me who depend on it for most of our driving.

Just look at how much better Autopilot has gotten over the years. When does all that improvement stop? When does it stop getting better? If anything, the rate of improvement for Autopilot is actually increasing. It’s clear to many people who are paying attention that the software will continue to improve iteratively until it’s good enough to be fully driverless. As that happens, Tesla is producing millions of additional cars that will run the software and shipping them to every corner of the Earth.

Watch the Drives

I have posted the entire 9.5 hours of driving to YouTube. If you’ve ever wanted to sit in traffic for ten hours on Autopilot, now you can do so virtually! This is easily the most boring video on YouTube today, but if you want to watch it you can. As with the other drives, I wanted to post it to document history and show people what is already possible today before Tesla has even released FSD Beta to all customers. If you don’t have ten hours to kill, there’s also a 10 minute timelapse version where every minute of video corresponds to roughly one hour of driving in the real world.

Here’s the timelapse of the drive:

And here’s nine and a half hours of raw footage:

There are a few frames cut out of the raw footage to protect the privacy of people who called me by not showing their names on the screen. I’m sure someone will accuse me of having a takeover during these dropped frames, but there were none. The missing frames are visible in the raw footage but all very short. If there is ever any doubt, I could always just show that part of the video with the name on the screen blurred out but I was feeling too lazy to do that on Sunday after 10 hours of driving.

This software will change the world, and few are paying attention

But soon, they will start to notice. Buckle up everyone. Electrifying all of America’s rides was never going to happen in a reasonable timeframe before, but with Autopilot we finally have a realistic shot at a plan to get everyone using electric cars for their day-to-day travel.

This isn’t about the future anymore. It’s already changing the world today. I would have lost my mind handling this stop-and-go traffic myself. The way that Autopilot was able to handle the whole drive made the whole situation a bit less stressful, and even today in a pre-release beta state, FSD is already indispensable to me. Just imagine what the first public release will be capable of. Just imagine what the release one year after that looks like. This is a massive revenue opportunity for Tesla short term, long before fully driverless Teslas are roaming the streets on their own.

Previous Drives

Here are two previous drives from Los Angeles to Silicon Valley.

Thanks for visiting and reading my stupid thoughts on this incredible transformation.